Improving Polymer Property Predictions with Geometric Information in Equivariant Graph Neural Networks

Polymers are versatile materials with a wide range of applications. The application of machine learning (ML) models in property prediction and inverse design are accelerating computational polymer screening and their automated design and optimization. The acceleration emerges from improved feature vector or graph representations which underly advanced ML models. While such representations can encode chemical information with regards to bonds, topology, and chemical environment, certain geometrical aspects are still lacking. Geometric features such as torsion angles between monomer units within polymer chains or the orientation of polymers chains with respect to each other determine, to a large extent, a polymer’s functional properties, including glass transition temperature and gas permeability.

In this contribution, we demonstrate how to leverage geometrical information through POS-EGNN [1]. POS-EGNN is a Position-based Equivariant Graph Neural Network for Materials Discovery in which each layer encodes many-body information representing a material’s atomistic geometry. Specifically, we have investigated effects related to the number of repeat units in polymer chains and to the number of chains in a representative volume for predicting a polymer’s glass transition temperature and gas permeabilities. We discuss how the inclusion of geometrical information affects the prediction accuracy of POS-EGNN in comparison with other ML architectures.

References [1] POS-EGNN is available at https://github.com/IBM/materials/tree/main/models/pos_egnn